Contamination MapTabular / Model
TabPFN v2
- Organisation
- Prior Labs / University of Freiburg
- Family
- TabPFN · version v2
- Released
- 2025-01-08 (Nature publication date ('Published: 08 January 2025') https://www.nature.com/articles/s41586-024-08328-6)
- Checkpoints
- Prior-Labs/TabPFN-v2-clf, Prior-Labs/TabPFN-v2-reg
- Weights
- Open · Prior Labs License (Apache 2.0 with additional attribution requirement)
- Links
- PaperModel cardCode
Training data disclosure: Full list
Level determined for 100% of benchmark cells.
- TestTest data seen0
- SeenSeen, split unclear0
- TrainTraining data seen0
- Related data seen0
- NoNot seen51
- ?Unknown0
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Other declared training data
-
TabPFN v2 synthetic prior (structural-causal-model generated datasets) · split: unknown
Evidence
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
www.nature.com · paper · Introduction · retrieved 2026-10-15 “we created a massive corpus of around 100 million synthetic datasets per model training, each with a unique causal structure, feature types and functional characteristics.”
www.nature.com · paper · Synthetic data based on causal models · retrieved 2026-10-15
Benchmark cells
TabArena
Not seen airfoil_self_noise · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Amazon_employee_access · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen anneal · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Another-Dataset-on-used-Fiat-500 · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen APSFailure · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen bank-marketing · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Bank_Customer_Churn · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Bioresponse · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen blood-transfusion-service-center · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen churn · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen coil2000_insurance_policies · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen concrete_compressive_strength · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen credit-g · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen credit_card_clients_default · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen customer_satisfaction_in_airline · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen diabetes · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Diabetes130US · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen diamonds · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen E-CommereShippingData · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Fitness_Club · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Food_Delivery_Time · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen GiveMeSomeCredit · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen hazelnut-spread-contaminant-detection · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen healthcare_insurance_expenses · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen heloc · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen hiva_agnostic · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen houses · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen HR_Analytics_Job_Change_of_Data_Scientists · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen in_vehicle_coupon_recommendation · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Is-this-a-good-customer · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen kddcup09_appetency · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen Marketing_Campaign · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen maternal_health_risk · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen miami_housing · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen MIC · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen NATICUSdroid · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen online_shoppers_intention · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen physiochemical_protein · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen polish_companies_bankruptcy · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen qsar-biodeg · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen QSAR-TID-11 · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen QSAR_fish_toxicity · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen SDSS17 · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen seismic-bumps · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen splice · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen students_dropout_and_academic_success · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen superconductivity · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen taiwanese_bankruptcy_prediction · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen website_phishing · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen wine_quality · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Not seen jm1 · absent from the complete list of training data
Evidence (2)
“By relying on synthetic data instead of large collections of public tabular data, we avoid common problems of foundational models, such as privacy and copyright infringements, contaminating our training data with test data”
“Before being applied to real-world datasets, the model is once pre-trained on millions of synthetic datasets representing different prediction tasks.”
Notes
Pre-trained only on synthetic data from an SCM-based prior. The HF classifier repo also holds alternative checkpoints (e.g. tabpfn-v2-classifier-v2_default.ckpt); the card does not say whether any of them involve real data (see open questions).
Sources
- Accurate predictions on small data with a tabular foundation model (Nature 2025) · paper · retrieved 2026-10-15
- Prior-Labs/TabPFN-v2-clf model card · model card · retrieved 2026-10-15
- Prior-Labs/TabPFN-v2-reg model card · model card · retrieved 2026-10-15